Adaptive Watermark Partitioning for High-Capacity Watermarks in Neural Networks: Balancing Robustness and Accuracy

This paper proposes an adaptive watermark partitioning framework to secure deep neural networks (DNNs) with high-capacity watermarks while preserving model performance. The framework dynamically allocates watermark bits based on class-specific characteristics, introducing two strategies: loss-based allocation, which prioritizes classes with lower classification loss for enhanced robustness, and uniform-based allocation which distributes watermark bits equally across all classes. To evaluate the trade-off between embedding capacity and model utility, we introduce the Accuracy-Capacity Tradeoff (ACT) metric. Extensive experiments on ResNet18 on CIFAR-10 and EfficientNetB0 on STL10 demonstrate that our method successfully embeds 1,024 to 2,048 bits with minimal bit error rate and accuracy degradation (<1%), achieving cryptographic-scale uniqueness (2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1024</sup> identifiers). The loss-based scheme reduces Bit Error Rate (BER) by 12-15% compared to the uniform scheme. These results underscore the framework's potential to enhance intellectual property protection for DNNs in real-world deployments, offering a scalable solution for balancing security and performance.